Molecular dynamics (MD) simulations are central to materials and drug discovery yet remain computationally demanding, particularly for free-energy perturbation (FEP) protocols and rare-event sampling. Existing sequence-based accelerators, especially Long Short-Term Memory (LSTM) models, often fail to capture long-range temporal structure and provide sufficient expressive capacity in noisy trajectories. Here, we introduce BiLSTMK-MD, a neural time-series learning method that constructs a causality-preserving surrogate for MD and FEP trajectories to reduce sampling requirements. The approach couples a sliding-window bidirectional LSTM encoder, which preserves long-range correlations, with an attention mechanism to enhance temporally informative frames, while a Kolmogorov-Arnold network output layer applies expressive nonlinear readout to decode the attention-refined representation into the final output. A two-stage, fANOVA-guided Bayesian optimization process searches for the optimal hyperparameter configuration for each system. Across four data sets, BiLSTMK-MD achieves mean absolute errors below 1.5 kcal mol-1 for window-resolved free-energy increments, reconstructs dihedral free-energy basins from 1-10% of trajectories, and maintains performance across systems. In long-trajectory regimes, it affords up to 400-fold acceleration for FEP and >700-fold speedup for rare-conformation sampling over conventional MD/FEP simulation. This neural time-series surrogate therefore provides a route to reducing sampling demands for free-energy estimation and rare-event characterization.
Mo et al. (Wed,) studied this question.